OK Mark,
Yes, the model does depend upon having accurate parameters obtained by calibration. The model will need to be modified if the cell is changed, but that is to be expected since it attempts to match the performance of the cell. I just began working on the EU cell and the results are pretty good so far. My first attempt was to use the calibration run on 12/7/2012 to define the quadratic values. They again were accurate to R^2=.9998 or so which is pretty good. With these a, b, c terms I used my model to predict the time domain response. The first run with with the power changing from .036 watts to 28.9 watts during the calibration run matched with an error of .5 degrees or so. I think the 0 power level gives the program a tough point to work with. Next I went from 28.8 watts to 38.6 watts for the second step of their run. Here the curve was beautiful as with the USA cell. The noise level was less than .25 volts with a sinusoidal addition again that dominated the noise. The period of the sine wave was roughly 1000 seconds. I would estimate that the sine wave was about equal to the average noise alone. I am very encouraged by these results. It will be most interesting when my simulation is applied to the systems with expected excess power. It should stand out very well against the calibration data. Dave -----Original Message----- From: MarkI-ZeroPoint <[email protected]> To: vortex-l <[email protected]> Sent: Tue, Dec 25, 2012 4:25 pm Subject: RE: [Vo]:Non Linear Model of Celani Device Thanks Dave! So one sigma is ~0.25 degsC, and that’s for several thousand points, so confidence level is high… No need for any other calcs at this time; just wanted to get an idea of the level of uncertainty. Your model and the noise level are tied to the experimental setup and process; if any changes are made to the setup, your model may no longer apply… but I’m sure you know all that! Hope the ones doing the tests understand all this… -Mark From: David Roberson [mailto:[email protected]] Sent: Tuesday, December 25, 2012 11:24 AM To: [email protected] Subject: Re: [Vo]:Non Linear Model of Celani Device Mark, I just let Excel run a standard deviation for all the points of the data series throughout the range of the experiment and obtained .24916 degrees C. This includes a time frame that begins at 0 seconds and continues to 9541 seconds. Each point is typically 2 to 3 seconds away from it's neighbors. The total number is 5508 data points for the standard deviation calculation. Do you wish for me to perform additional tests upon the output? Dave -----Original Message----- From: David Roberson <[email protected]> To: vortex-l <[email protected]> Sent: Tue, Dec 25, 2012 2:08 pm Subject: Re: [Vo]:Non Linear Model of Celani Device Mark, I can give you a hint as to how well the model matches the actual real life data. I have plotted a curve of the difference between the actual data and my model prediction. The difference looks like random noise that is more or less evenly distributed about 0 volts throughout the entire power input to temperature output transition. This includes the case I analyzed beginning at 48.2 watts and ending with 82.7 watts. I see no evidence of any curvature associated with the error between my simulation and the real data. There is a small, almost sinusoidal, signal hidden deeply within the noise that continues throughout the entire time frame which in this case is 9541 seconds long. The total noise peaks tend to be in the vicinity of .5 degrees C while the average of the flat noise is more in line with .2 degrees C. Perhaps I should make a plot of the output and send it for you to review. It is pretty impressive to see consistent noise when the large time domain transition signal is balanced out. My mention of the possible excess power is based upon my having to include an additional 1 watt of input power for my model to achieve the perfect match. It is quite obvious that the extra power is required for the curve to fit so perfectly. The data I used was from 11/30/2012 at 2200 hours according to my download from the MFMP replication site. I used the history points for my curve fitting and analysis. I fitted the transition between the two power levels shown above. I just took a look at the small noisy sinusoidal signal hidden within the noise and it appears to be in the ballpark of 2000 seconds in period. Maybe this corresponds to the cycle time for the heating system. I guess I can attempt an RMS noise measurement which will be next on my list. The small sinusoidal interference will color that result a bit. I will report the results of the test when completed. Dave -----Original Message----- From: MarkI-ZeroPoint <[email protected]> To: vortex-l <[email protected]> Sent: Tue, Dec 25, 2012 12:18 pm Subject: RE: [Vo]:Non Linear Model of Celani Device Dave: Can you perform some stats on the model vs reality and give us the std deviation? -Mark From: David Roberson [mailto:[email protected]] Sent: Tuesday, December 25, 2012 9:15 AM To: [email protected] Subject: Re: [Vo]:Non Linear Model of Celani Device During the night Santa brought me a gift! A thought occurred to me that there is a very good explanation for the 30 to 40 second time constant exponential waveform that I have been seeking. In order to get the best curve fit to the exact solution of the differential equation I have been forced to modify the constant of integration slightly away from the ideal value as determined by steady state measurements. This seemed strange, but now I realize that it is required to compensate for the displacement of the rising edge due to the above delay. It is necessary to add back the initial plug of energy lost when the best differential equation solution is followed. This ideal solution for the best overall data match must start at a value that is below the actual temperature of the cell at t=0 in order to accommodate the delayed behavior. The addition of this missing energy is exactly the amount required! So now I can say with confidence that there exists a delay mechanism which retards the reading of the temperature at the outer glass surface. This delay is in addition to the ideal non linear differential equation solution time domain response which is discussed below. So, another way to envision the effect is to realize that it takes 30 to 40 seconds before the addition of heat applied to the cell is registered at that test point. An exponential smoothing (filtering) factor is applied. My suspicion is that the extra pulse of heat must be distributed within the gas and then result in a temperature reading at the outer glass monitor after heating the envelop. The heating of the other structure elements may also be involved in the overall action. A careful review of the waveform hints that the test might be demonstrating an excess power of about 1 watt during the experiment that supplied the data. This is a small amount of excess power and only additional, careful analysis would enable me to be sure. At least it is in the right direction! My virtually perfect curve fit to the data tends to support this conclusion. Merry Christmas! Dave -----Original Message----- From: David Roberson <[email protected]> To: vortex-l <[email protected]> Sent: Tue, Dec 25, 2012 2:13 am Subject: [Vo]:Non Linear Model of Celani Device The data has been flooding in from the MFMP and I have been seeking a time domain model of the system behavior when power to the Celani replication device is modified. Most of my effort has been exerted by analyzing the rising edge of the time domain waveform when the drive power is stepped up by a significant amount. The temperature follows a certain path as it ramps up to the value required to balance the input and output power of the cell. We have been fortunate in this particular case to find that the relationship between temperature and input power is well behaved and follows a second order curve to a remarkable degree. It is not uncommon to see a curve fit with R^2=.9999 or better in many independent test runs. I initially was expecting to see a power series that included a forth order term of significance due to the S-B radiation equation. This has not ever been dominate in any test and I still am trying to understand why this is true. For the time being I will accept this gift happily. A quick glance at the shape of the rising edge of the temperature curve suggests that it follows an exponential. I thus began my model by making that assumption and got fairly reasonable results. It was always evident that my curve fit contained holes, but a couple of degrees of error did not seem too excessive at that time. Being a perfectionist, I decided to improve the situation and to determine how well a model could match the real life test. I very soon added a second exponential to the mix and noticed that the fit improved remarkably. Also, I noticed that the second real frequency was close to the second harmonic of the first one determined by my earlier work. A light went off inside my head and I realized that this would be expected since the non linearity is mainly of second order in the relationship between variables. Now, I saw that the accuracy of my model was becoming very acceptable. There remained a short period of time at the initial power increase where the fit was not as good as I hoped. To fix this problem I added another exponential with an associated time constant of about 40 seconds. With this model, I could obtain an excellent match between my simulation and the real world data. I could have left it in this state, but it is hard to accept imperfection. To pursue the matter further I used a LTSpice model of the system. I guessed correctly in my first try with the model and was rewarded with a well behaved simulation that included the second order distortion effects. This model was used for a significant time as it matched the real world waveforms everywhere except for the initial short period that required another time constant to fix. Looking at my spice model gave me an interesting idea. I used a capacitor to represent storage of the incoming energy and the node it is connected to reads expected time domain temperature for the outside glass sensor. In parallel with the storage capacitor is a pair of current sources, one representing power applied to the cell, the other power being taken away by the various paths. The draining current source appears as a parallel conductance who's value depends upon the voltage at the temperature node. I, of course, was seeking verification of the time constant associated with the exponential rise waveforms and attempted to use the effective conductance value in parallel with my storage capacitor for a quick check. This lead to the non linear differential equation definition that works so well. It occurred to me that my model could be expressed in the form of a non linear differential equation with a little manipulation of the shape. Basically you have a parallel capacitor being driven by a current source that is paralleled by a non linear conductance. The non linear conductance is neatly defined by the second order equation derived from the calibration runs for the Celani cell. Now, all I had to do was to solve the non linear differential equation that I constructed and insert the initial conditions to define the temperature and power over any time frame. My first thought was yipes! I consulted our favorite source wikipedia to find the solution to unusual integrals. The one I needed to solve was in the form of: Integral dx/(a*x^2+b*x+c) with initial condition of the temperature of the steady state value just prior to the application of an increase in power. I transformed the time scale so that time = 0 was with this application of extra power. It turns out that there is an exact solution to such an equation which you can look up at your convenience to save time and space here. I had to perform some interesting series adjustments to get the curve within the desired temperature band, and I was a bit rusty at first. Finally, a perfect curve was being generated that matched the time domain data extremely well except for that nagging time region at the very start. I continue to have to include an additional exponentially rising pulse function with a time constant around 40 seconds at the application of the extra drive to get virtually perfect tracking to the real world data. Next, I included another drive current waveform of this nature to my spice model and it tweaks the start of the rising edge a tiny amount much like a delay. I am still seeking a good explanation for the necessity of this extra pulse source and I wonder if it can be traced to the IR effects or some other relatively large time constant such as the glass heating. The nature of the extra leading edge drive pulse can be described as a signal that begins at a certain level and decays exponentially to zero with a time constant of 40 seconds. The effective DC component of the waveform is taken out by the action of the non linear conductance. One interesting observation is that the calibration determined a,b, and c that constitute coefficients of the second order equation defining Power versus Outside Glass Temperature along with the input power uniquely determine the steady state temperature of the device. These four variables define operation over the entire range of input powers. My model also includes a capacitor that acts as the energy storage stand in. One good temperature rising transition allows me to choose the correct capacitor to enter into the model. The additional short time exponential must be determined by curve fitting within a short initial period typically 100 seconds. I have found this exercise interesting and educational. If a good explanation for that initial power pulse is obtained I can relax and fool with the incoming data. I am hoping that my contributions will enable us to discover any excess power that may occur by its signature outside of the normal that I now model and observe. Dave
